the industry hasn't yet internalized how much serverless tool calling will change agent economics watch me run 1000 agents in parallel on a single archil sandbox, with tool calls going to actual Linux microVMs -- all for 30x less than a traditional sandbox it doesn't matter how cheap or... expensive the underlying cpus are if you don't waste them while inference is running the cpu capacity crisis will be solved by running agents on right-sized hardwareshow more

so cool

Seems very cool... but I don't get it... let me try: Ok so architecture is 1000x agents on one box (makes sense) -> calling into microVM workspaces somewhere else? I'm confused of the savings here? Is this not just classic use of overpacking firecracker? is this compared to running an agent _IN_ a sandbox? Or just compared to margin prices on off the shelf sandboxes?

yes, this is compared to agent-in-sandbox, but similar order of magnitude for agent-outside-of-sandbox if you're still launching that sandbox on the first request and running through the end of the session going to do a longer post on what this means today

let me know if this answers

thanks

Paying for idle CPU while the model is thinking is a cost nobody talks about, curious how much latency a microVM per tool call adds compared to a warm sandbox

compared to a running sandbox, you pay about 40ms more

harness outside the sandbox is the way

What's your setup for creating these animated charts?

these days i mostly tell the chatgpt app to use opus 5.5 for design by calling the "claude" CLI and then... idk it just kind of happens took an annoyingly long time to get it in the right (even now i don't love it) shape

Gotcha I have the same issue that I spend more time fiddling with it than I want Thanks for sharing

very cool

what's the cold start on a microVM spun up per tool call? does it still beat a warm traditional sandbox when the call takes under a second?

it's about 70ms

Cheap agent runs matter only when the task is safe and repeatable. The real win is more work without more ops work.

are you saying that each agent is making an independent tool call in the same vm?

very cool

This is the right shape of agent infra: keep inference scarce and burst tool work into cheap isolated sandboxes. The missing primitive is a run receipt per tool call: sandbox image, policy, wall time, exit code, artifact hash, cost. Then schedulers can price behavior, not vibes.

Cheap parallel tool calls change the failure math too. At 1000 agents, a 1% flaky tool is 10 broken traces per run, so retries need to show up in the trace, not get hidden inside the sandbox.

cool 👍

The part that decides whether this holds for coding agents is state between calls. A build that needs the deps and cache the last tool call left behind pays a big rehydration cost on a cold microVM, so keeping the working tree warm while the CPU is released is the real trick.

most of the agent container time is just idle, waiting for model response, and we still pay full Fargate for it. When one tool call needs files written by the previous one, does Archil keep that state for the next microVM or each call starts fresh?

microVM is one of main technologies enabling I drawing attention to one specific problem solved there, allowing agents to interact with processes, no just launching them.

Does this also work with agents that need a filesystem f.e. for development work? I'm looking for a cheap elastic alternative right now for non eu users

The isolation per call is real. But a fresh microVM per tool call doesn't change what the tool can reach: the credentials the call carries still define that. Cheap sandboxes scale reach as easily as they scale agents.

did opus make this animation

Let’s have a conversation. dm me